LECTURE 11
Introduction to Econometrics
Autocorrelation
November 29, 2016
1 / 24
ON PREVIOUS LECTURES
I We discussed the specification of a regression equation
I Specification consists of choosing:
1. correct independent variables2. correct functional form3. correct form of the stochastic error term
I We talked about the choice of independent variables andtheir functional form
I We started to talk about the form of the error term - wediscussed heteroskedasticity
2 / 24
ON TODAY’S LECTURE
I We will finish the discussion of the form of the error termby talking about autocorrelation (or serial correlation)
I We will learn
I what is the nature of the problem
I what are its consequences
I how it is diagnosed
I what are the remedies available
3 / 24
NATURE OF AUTOCORRELATION
I Observations of the error term are correlated with eachother
Cov(εi, εj) 6= 0 , i 6= j
I Violation of one of the classical assumptions
I Can exist in any data in which the order of theobservations has some meaning - most frequently intime-series data
I Particular form of autocorrelation - AR(p) process:
εt = ρ1εt−1 + ρ2εt−2 + . . .+ ρpεt−p + ut
I ut is a classical (not autocorrelated) error termI ρk are autocorrelation coefficients (between -1 and 1)
4 / 24
EXAMPLES OF PURE AUTOCORRELATION
I Distribution of the error term has autocorrelation nature
I First order autocorrelation
εt = ρ1εt−1 + ut
I positive serial correlation: ρ1 is positiveI negative serial correlation: ρ1 is negativeI no serial correlation: ρ1 is zeroI positive autocorrelation very common in time series dataI e.g.: a shock to GDP persists for more than one period
I Seasonal autocorrelation (in quarterly data)
εt = ρ4εt−4 + ut
5 / 24
EXAMPLES OF IMPURE AUTOCORRELATIONI Autocorrelation caused by specification error in the
equation:I omitted variableI incorrect functional form
I How can misspecification cause autocorrelation in theerror term?
I Recall that the error term includes the omitted variables,nonlinearities, measurement error, and the classical errorterm.
I If we omit a serially correlated variable, it is included in theerror term, causing the autocorrelation problem.
I Impure autocorrelation can be corrected by better choice ofspecification (as opposed to pure autocorrelation).
6 / 24
AUTOCORRELATION
X
Y
7 / 24
CONSEQUENCES OF AUTOCORRELATION
1. Estimated coefficients (β̂) remain unbiased and consistent
2. Standard errors of coefficients (s.e.(β̂)) are biased(inference is incorrect)
I serially correlated error term causes the dependent variableto fluctuate in a way that the OLS estimation procedureattributes to the independent variable
I Serial correlation typically makes OLS underestimate thestandard errors of coefficients
I therefore we find t scores that are incorrectly too high
⇒ The same consequences as for the heteroskedasticity
8 / 24
DURBIN-WATSON TEST FOR AUTOCORRELATION
I Used to determine if there is a first-order serial correlationby examining the residuals of the equation
I Assumptions (criteria for using this test):
I The regression includes the intercept
I If autocorrelation is present, it is of AR(1) type:
εt = ρεt−1 + ut
I The regression does not include a lagged dependentvariable
9 / 24
DURBIN-WATSON TEST FOR AUTOCORRELATION
I Durbin-Watson d statistic (for T observations):
d =
T∑t=2
(et − et−1)2
T∑t=1
e2t
≈ 2(1− ρ̂)
where ρ̂ is the autocorrelation coefficient
I Values:
1. Extreme positive serial correlation: d ≈ 02. Extreme negative serial correlation: d ≈ 43. No serial correlation: d ≈ 2
10 / 24
USING THE DURBIN-WATSON TEST
1. Estimate the equation by OLS, save the residuals
2. Calculate the d statistic
3. Determine the sample size T and the number ofexplanatory variables (excluding the intercept!) k′
4. Find the upper critical value dU and the lower criticalvalue dL for T and k′ in statistical tables
5. Evaluate the test as one-sided or two-sided (see next slides)
11 / 24
ONE-SIDED DURBIN-WATSON TEST
I For cases when we consider only positive serial correlationas an option
I Hypothesis:
H0 : ρ ≤ 0 (no positive serial correlation)HA : ρ > 0 (positive serial correlation)
I Decision rule:I if d < dL reject H0
I if d > dU do not reject H0
I if dL ≤ d ≤ dU inconclusive
12 / 24
DURBIN-WATSON CRITICAL VALUES FOR ONE-SIDED
TEST
13 / 24
TWO-SIDED DURBIN-WATSON TEST
I For cases when we consider both signs of serial correlation
I Hypothesis:
H0 : ρ = 0 (no serial correlation)HA : ρ 6= 0 (serial correlation)
I Decision rule:
I if d < dL reject H0
I if d > 4− dL reject H0
I if d > dU do not reject H0
I if d < 4− dU do not reject H0
I otherwise inconclusive
14 / 24
DURBIN-WATSON CRITICAL VALUES FOR TWO-SIDED
TEST
15 / 24
EXAMPLE
I Estimating housing prices in the UK
I Quarterly time series data on prices of a representativehouse in UK (in £)
I Explanatory variable: GDP (in billions of £)
I Time span: 1975 Q1 - 2011 Q2
I All series are seasonally adjusted and in real prices (i.e.adjusted for inflation)
16 / 24
EXAMPLE
60000
80000
100000
120000
140000
160000
180000
200000
220000
1975 1980 1985 1990 1995 2000 2005 2010
Pri
ce o
f re
pre
senta
tive h
ouse
17 / 24
EXAMPLE
18 / 24
EXAMPLE
I We test for positive serial correlation:
H0 : ρ ≤ 0 (no positive serial correlation)HA : ρ > 0 (positive serial correlation)
I One-sided DW critical values at 95% confidence forT = 146 and k′ = 1 are:
dL = 1.72 and dU = 1.74
I Decision rule:I if d < 1.72 reject H0
I if d > 1.74 do not reject H0
I if 1.72 ≤ d ≤ 1.74 inconclusive
I Since d = 0.02 < 1.72, we reject the null hypothesis of nopositive serial correlation
19 / 24
ALTERNATIVE APPROACH TO AUTOCORRELATION
TESTING
I Suppose we suspect the stochastic error term to be AR(p)
εt = ρ1εt−1 + ρ2εt−2 + . . .+ ρpεt−p + ut
I Since OLS is consistent even under autocorrelation, theresiduals are consistent estimates of the stochastic errorterm
I Hence, it is sufficient to:
1. Estimate the original model by OLS, save the residuals et
2. Regress et = ρ1et−1 + ρ2et−2 + . . .+ ρpet−p + ut
3. Test if ρ1 = ρ2 = . . . = ρp = 0 using the standard F-test
20 / 24
BACK TO EXAMPLE
21 / 24
BACK TO EXAMPLE
22 / 24
REMEDY: WHITE ROBUST STANDARD ERRORS
I Note that autocorrelation does not lead to inconsistentestimates, only to incorrect inference - similar toheteroskedasticity problem
I We can keep the estimated coefficients, and only adjust thestandard errors
I The White robust standard errors solve not onlyheteroskedasticity, but also serial correlation
I Note also that all derived results hold if the assumptionCov(x, ε) = 0 is not violated
I First make sure the specification of the model is correct,only then try to correct for the form of an error term!
23 / 24
SUMMARY
I Autocorrelation does not lead to inconsistent estimates,but it makes the inference wrong (estimated coefficientsare correct, but their standard errors are not)
I It can be diagnosed using
I Durbin-Watson testI Analysis of residuals
I It can be remedied by
I White robust standard errors
I Readings:I Studenmund, Chapter 9I Wooldridge, Chapter 12
24 / 24